Skip to content

claude-code.md

Latest commit

 

History

History
61 lines (42 loc) · 1.5 KB

claude-code.md

File metadata and controls

61 lines (42 loc) · 1.5 KB

Claude Code Integration

engram integrates with Claude Code via PreToolUse/PostToolUse hooks that intercept Read/Edit/Write calls and inject structural context automatically.

Prerequisites

# Index your project
engram init .

# Install hooks into Claude Code settings
engram install-hook

install-hook writes to .claude/settings.json (project scope) or ~/.claude/settings.json (user scope, pass --scope user).

What happens at runtime

Every Read call is intercepted. Instead of passing the full file to Claude, engram returns a structural summary: node list, key relationships, and confidence scores. Full file content is still available on demand — just re-read with explicit offset/limit params.

The hook exits with deny + reason to inject context inline, which means zero latency overhead for Claude Code (synchronous hook path).

HUD

The status line shows live hit rate and estimated token savings:

engram ◆ 47 hits · ~12K tokens saved

Configure the label in .engram/providers.json.

Dashboard

engram dashboard

Opens a live terminal view showing recent intercepts, node coverage, and per-file hit rates.

Provider configuration

.engram/providers.json controls which context providers are active and their budget limits:

{
  "structure": { "enabled": true, "budget": 800 },
  "mistakes":  { "enabled": true, "budget": 200 },
  "decisions": { "enabled": true, "budget": 200 }
}

Run engram stats -p . for a summary of current graph state.